AI Is Writing More Code Than Human Programmers — What This Means for Software Jobs

GitHub reports AI now generates over 50% of new code on its platform. Here is what this means for software engineering jobs, salaries, and which skills remain human-exclusive.

GitHub's data — released in early 2026 — that AI assistance now generates more than 50 percent of new code written on its platform represents a milestone that the software engineering industry has been tracking with a combination of professional anxiety and technological fascination. The specific figure requires context to be meaningful: GitHub Copilot and similar AI coding assistants generate code that human programmers then accept, modify, or reject. The 50 percent figure reflects AI generation, not AI-independent autonomous coding — humans remain in the loop for every line that ends up in production.

The impact on software engineering employment is not the dramatic job elimination that the 50 percent figure might suggest but a specific productivity restructuring. Individual developers with AI assistance are completing tasks in approximately 55-70 percent of the time that unassisted development previously required (McKinsey's study of AI coding tools' productivity effect). This means that the same software output can be produced with fewer developer-hours — which creates ambiguous labour market dynamics: existing developer headcount can produce more output (good for productivity), or existing software needs can be met with fewer developers (potential employment contraction).

The specific skills that remain most valuable for human programmers in the AI-assisted era: architectural decision-making (AI generates code well when the system design is clear; it struggles with novel architectural problems whose solutions require genuine creativity and systems thinking); security review (AI-generated code reliably produces specific categories of security vulnerability including SQL injection and cross-site scripting that human review must identify); and stakeholder requirements translation (the natural language to technical specification conversion that determines what software is built is poorly suited to AI automation).

For the education and training pipeline: computer science education is restructuring around the assumption that AI will handle boilerplate code generation, shifting curriculum emphasis toward system design, security, and the human skills that AI augments rather than replaces. Whether this restructuring is happening fast enough relative to the pace of AI capability development is actively debated.